增强急诊外科决策:机器学习驱动风险预测模型的系统评价
该研究系统评价了急诊外科(急诊普通外科、外科重症、创伤)中机器学习驱动风险预测模型的建模技术、预测特征和结局。共纳入52项研究,57.7%聚焦创伤人群,76.8%使用登记数据,仅5项在模型开发后进行了外部验证。最常见结局为死亡(59.6%)、并发症(30.8%)和分诊/严重度(15.4%),无患者中心/报告结局。机器学习模型在11项比较研究中普遍优于传统风险评分。结论认为性能有前景,但方法学严谨性、真实世界验证不足及特征、结局和算法异质性大,挑战系统化采纳与监管。
要点
- 52项研究中57.7%聚焦创伤人群,76.8%使用登记数据,仅5项在开发后进行了外部验证。
- 死亡(59.6%)、并发症(30.8%)、分诊/严重度(15.4%)为最常见预测结局,患者中心/报告结局缺如。
- 常用算法为逻辑回归(44.2%)、随机森林(34.6%)和决策树(26.9%),支持向量机表现最高。
- 模型性能通常以受试者工作特征曲线下面积评估(88.5%),11项比较研究中机器学习模型普遍优于传统风险评分。
- 影像和术中数据利用不足,自然语言处理仅用于4项研究。
适用人群与场景
- 急诊外科人群,包括急诊普通外科、外科重症和创伤。
- 52项研究中57.7%聚焦创伤人群。
研究设计与关键结果
- 系统评价纳入52项研究,57.7%聚焦创伤人群;76.8%使用登记数据,28.8%使用电子健康记录,仅5项在模型开发后进行外部验证。
- 常用算法包括逻辑回归(44.2%)、随机森林(34.6%)、决策树(26.9%);支持向量机表现最高。
- 最常见预测结局为死亡(59.6%)、并发症(30.8%)、分诊/严重度(15.4%);患者中心/报告结局缺如。
- 常用特征包括人口学、生理评分和生命体征,影像和术中数据利用不足;自然语言处理用于4项研究。
- 模型性能通常以受试者工作特征曲线下面积评估(88.5%);11项比较研究中机器学习模型普遍优于传统风险评分。
证据类型、等级与局限
- 证据类型:系统评价(无Meta分析);证据等级:Level IV。
- 局限:方法学严谨性有限、真实世界验证不足,特征、结局和算法存在显著异质性。
对床旁实践的改变
- 机器学习模型在急诊外科风险预测中性能有前景,但方法学严谨性、真实世界验证不足及特征、结局和算法异质性大,挑战系统化采纳与监管。
- 深入理解急诊外科决策工作流程及其实施后影响,可能有助于确保机器学习风险预测的临床相关性、无缝和安全整合。
提炼说明
- 本条依据论文摘要整理;摘要未涉及「诊断标准与关键阈值」「处置流程与药物剂量」「红旗征与不能漏的情况」「指南或共识出处」,故未列出这些栏目;全文请点「阅读原文」。
英文原文(摘要)· 点击展开
BACKGROUND: Acute care surgery (ACS) involves rapid, high-stakes decisions with limited opportunity for preoperative planning. While machine learning (ML) may improve risk prediction and decision making in this setting, its development, validation, and implementation in ACS remain understudied. We therefore evaluated the techniques, predictor features, and outcomes used in ML-driven risk prediction models in ACS and generated recommendations to inform future research and support clinically meaningful implementation.
METHODS: A systematic review of ML-driven predictive models in ACS (emergency general surgery, surgical critical care, trauma) was conducted. Models were analyzed by predictor features, outcomes, algorithms, and performance. The best-performing models for the most commonly predicted outcome were identified.
RESULTS: Of 52 studies, 57.7% focused on trauma populations. Most models used registry data (76.8%), fewer used electronic health records (28.8%), and only five studies performed external validation after model development. Common algorithms included logistic regression (44.2%), random forest (34.6%), and decision trees (26.9%). Mortality (59.6%), complications (30.8%), and triage/severity (15.4%) were the most frequent outcomes; patient-centered/reported outcomes were absent. Features commonly included demographics, physiologic scores, and vital signs, while imaging and intraoperative data were underused. Natural language processing was used in four studies. Model performance was typically assessed using area under the receiver operating characteristic curve (88.5%), with support vector machines demonstrating the highest performance. Machine learning models generally outperformed conventional risk scores among 11 comparative studies.
CONCLUSION: Machine learning-driven predictive models in ACS show promising performance but are constrained by limited methodological rigor, real-world validation, and substantial heterogeneity in features, outcomes, and algorithms, challenging systematic adoption and oversight. A grounded understanding of ACS decision making workflows and their postimplementation impact may ensure clinically relevant, seamless, and safe integration of ML-based risk prediction.
LEVEL OF EVIDENCE: Systematic Review Without Meta-analysis; Level IV.
参考来源
- Lee AH, Chan MK, Narayanan D, et al. Augmenting decision making in acute care surgery: A systematic review of machine learning-driven risk prediction models. J Trauma Acute Care Surg. 2025. PMID: 41196221. doi:10.1097/TA.0000000000004805
供执业医师抢救与诊疗参考,不替代临床判断;剂量请结合体重、肝肾功能和本院药品规格复核,以指南原文为准。